Implement window_size for decomposed forecasting in TimesFM 2.5

[philyuchkoff]

Summary

Implements the window_size parameter in ForecastConfig that was declared but never implemented (TODO(siriuz42)).

Decomposed forecasting was originally present in TimesFM 1.0 but was lost during the 2.5 migration. This PR restores the feature.

How it works

When window_size > 0 is set in ForecastConfig:

1. Each input time series is decomposed into a trend (moving average) and a residual component via moving_average()
2. Both components are fed to the model as independent inputs in the same batch
3. Output forecasts are summed element-wise: forecast(trend) + forecast(residual)

This improves long-range forecast quality by letting the model separately handle slow trends and fast residuals.

Changes

• src/timesfm/timesfm_2p5/timesfm_2p5_base.py: Added moving_average(), modified forecast(), blocked forecast_with_covariates() with explicit error when window_size > 0
• src/timesfm/configs.py: Replaced TODO(siriuz42) with full description
• tests/test_base_utils.py: 9 unit tests for moving_average() (all pass)
• tests/test_model_loading.py: Integration tests for forecast() with window_size and error case with covariates

Edge cases

• window_size = 0 (default): backwards-compatible, no behavior change
• window_size >= len(ts): uses actual array length, decomposition degenerates
• window_size = 1: trend equals original series, residual is zero
• Empty array
• forecast_with_covariates() raises ValueError with clear message

See docs/plans/2026-09-03-window-size-design.md for full design rationale and algorithm.

[sylvesterkaczmarek]

This drops the existing uncompiled-model guard. When forecast_config is None, the first condition is false and the elif not self.forecast_config.return_backcast dereferences None, so forecast_with_covariates() now raises AttributeError instead of the documented “Model is not compiled” ValueError. Please keep the compile check first and add an uncompiled regression.

[philyuchkoff]

@sylvesterkaczmarek
Thanks!
Fixed: restored the forecast_config is None check as the first guard, then the window_size check as the second. Both now raise ValueError with the expected messages.

[sylvesterkaczmarek]

Hi, decomposition now runs before the existing max_context truncation. For long inputs, the retained trend/residual values can therefore be influenced by samples outside the configured model context; the v1 path truncated first and then applied the moving average. Could the input be clipped to max_context before moving_average()?

[philyuchkoff]

@sylvesterkaczmarek
Thanks!
Fixed: input is now truncated to max_context before moving_average(), matching the v1 path order.

[sylvesterkaczmarek]

Re-checked both issues I raised. The uncompiled-model guard is again evaluated before window_size, preserving the documented ValueError, and long inputs are clipped to max_context before moving-average decomposition, matching the v1 ordering. Both findings are resolved.